Research On Lung Nodule Segmentation And Benign And Malignant Classification Algorithm Based On Deep Learning

Deep learning technology is increasingly widely used in medical image analysis,among which lung nodule segmentation and benign and malignant classification are key issues in the diagnosis and treatment of lung diseases.The traditional lung nodule segmentation method is affected by problems such as changeable nodule shape and blurred boundaries,resulting in unsatisfactory segmentation accuracy.The method based on deep learning has the advantage of extracting highquality image features and has better sensitivity to the size,position and edge shape of the image,while the current lung nodule segmentation and classification algorithms have different performance and low accuracy,especially for smaller lung nodules,which may cause some false detection and missed detection.Therefore,based on the LUNA16 dataset,this thesis proposes a V-Net based lung nodule segmentation algorithm and a Swin Transformer-based lung nodule benign and malignant classification algorithm.The experimental results show that the proposed model has significant superiority in image segmentation and classification of lung nodules.The main research contents are as follows:First,the V-Net network has insufficient image feature extraction and low feature fusion efficiency,resulting in its insensitivity to small nodules and poor segmentation effect.Therefore,this thesis proposes to add a hopping connection module based on residual mechanism between encoder and decoder,and the results show that the Dice coefficient of the improved PCV-Net network segmentation is increased by about 2%.Then,on the basic model of V-Net,this thesis proposes an SGV-Net segmentation model formed by fusing channel attention compression excitation module and attention guidance filter module,which is used to enhance the effective information of the region of interest and suppress irrelevant image background information,which enhances the network feature extraction ability when the overall complexity of the network is not high,and the improved model obtains more high-resolution information,which can restore the detailed information in the original image more perfectly.Experimental comparison shows that the Dice coefficient of SGV-Net is increased to 0.8452,which is about 5% higher than that of the benchmark V-Net model,and the segmentation results are more accurate.Second,in view of the current situation of low accuracy and poor robustness of benign and malignant classification algorithms for lung nodules,this thesis proposes a benign and malignant classification algorithm for lung nodules based on Swin Transformer.Based on the LUNA16 dataset,the algorithm is compared with common classification networks such as VGG,Res Net and Google Net,and the experimental results show that the accuracy of the classification results of the model reaches 94.53%,and the AUC value reaches 97.44%.Compared with other algorithms,the accuracy,sensitivity,specificity and AUC of the Swin Transformer model have significant advantages.